Building multi-agent systems often creates tangled toolsets and unmaintainable prompts. A new production-ready approach uses specialized AgentCore Runtime apps for distinct domains like HR and IT. Each agent handles its own retrieval from a dedicated knowledge base, exposing an A2A endpoint. This keeps responsibilities clean and deployments independent.

An orchestrator agent routes queries using tools like `ask_hr_agent` and `ask_it_agent`. Instead of hard-coded logic, the orchestrator’s LLM decides which agent to call based on the query. It synthesizes a final answer from the responses. This model-driven routing replaces brittle, manual decision trees with flexible, intelligent delegation.

Configuration is entirely environment-based, using variables like `KNOWLEDGE_BASE_ID` and `HR_AGENT_URL`. Handlers employ lazy initialization to optimize container performance, a key detail for production workloads. Failure modes are handled gracefully, allowing partial answers if one agent fails. This architecture scales easily—adding a Finance or Legal agent simply means repeating the pattern.

The result is essentially microservices for reasoning agents. Each component is small, focused, and communicates via a stable interface. This separation prevents knowledge leakage between domains and simplifies evaluation. Developers gain a clear evolution path from single RAG demos to robust, multi-domain systems that can run in real enterprise environments.